> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kelasa.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Workflows

> Deterministic, versioned automations built from typed steps

A **workflow** is a deterministic automation: a graph of typed steps that runs
the same way every time. Workflows are the reliable backbone of Kelasa — when
you need something to happen exactly, use a workflow; when you need judgment,
put an [agent step](#agent-steps) inside one or reach for a
[skill](/concepts/skills).

## Building

Describe what you want in the **AI builder** and it drafts the workflow as an
editable diagram — steps, branches, and connections you can inspect and refine
conversationally. You always see exactly what will run before it runs.

## Versions

Workflows are versioned. Editing never changes what's live: you save a new
version, review it, and publish deliberately. Published versions are what
triggers and the assistant execute.

## Runs

Every execution is recorded as a **run** with a live, step-by-step timeline —
watch it progress in real time, inspect each step's input and output, and see
exactly where and why something failed. Use **dry-run** to execute a workflow
without side effects while you're testing.

## Approvals

Any step can require an approval. The run pauses until someone with the right
role approves or denies — in the web app or from a Slack card. See
[Approvals](/guides/approvals).

## Agent steps

A workflow can include an **agent step**: a bounded AI loop that works inside
the run with the tools you granted it. The workflow stays deterministic around
it — the agent handles the part that needs judgment, and the run timeline
records what it did.

## Starting workflows

Workflows start from the assistant (on request), from
[triggers](/concepts/triggers) (events and schedules), or manually from the
workflow's page.